Are Online IQ Tests Accurate? What They Can and Cannot Tell You

An online IQ test is not accurate in the sense people want, and the reason has very little to do with whether its questions are any good. Accuracy in testing is a property of the entire procedure — the conditions, the comparison group, the administration and the interval between sittings. An unsupervised web test cannot control any of those, so its output has no defined meaning even when the items themselves are competently written.

That is a different complaint from the one usually made. This article is about delivery, not about the honest limits of a well-administered instrument, which are covered separately in the limitations of IQ testing.

Accuracy is a property of the procedure, not the questions

ETS defines reliable scores as those that are consistent across different occasions of testing, across different editions of a test, and across different raters scoring the same responses. Everything in that definition is about circumstances.

A normed instrument exists as a package: a fixed item set, a scripted administration, a trained administrator, a documented standardisation sample, and published evidence about consistency. Remove the packaging and the items are just questions. Copying the questions into a web form reproduces the least important part — as our guide to what IQ tests actually measure sets out, the items are the visible surface of a much larger apparatus.

The four things an unsupervised test cannot do

It cannot fix the conditions. The norms were collected under one set of conditions. At home there is no way to hold timing, interruptions, note-taking, a second screen or a second attempt constant. A score produced under unknown conditions cannot be compared with one produced under known conditions.

It cannot document the comparison group. A percentile means something only against a stated sample. Free web tests are typically ranking you against the people who chose to take that test on that site, which is self-selected in ways nobody has characterised. A rank against an undescribed group is not a rank.

It cannot observe. A live administrator does more than police cheating. They note whether the person understood the task, whether attention lapsed, whether an answer was a guess or a considered response — context that determines whether the numbers are even usable.

It cannot enforce a retest interval. Repeat exposure produces real score gains, discussed below, and nothing stops a web test being taken repeatedly until a pleasing number appears.

What the proctoring research actually found

It would be convenient if supervision were a single switch. The evidence is more interesting than that.

Scherrer and colleagues, writing in the Journal of Intelligence in 2025, compared three conditions across 226 participants in Germany: webcam-proctored testing, unproctored testing by random assignment, and unproctored testing that participants chose. The setting was low-stakes research rather than selection or diagnosis.

They found no significant overall performance difference by proctoring status. Proctored participants did slightly better on reasoning and divergent thinking, and slightly worse on memory tasks. Memory tasks also showed measurement invariance violations across conditions — meaning the tasks were not behaving identically in each setting, which is a measurement problem rather than a score problem. Observable cheating in the proctored sessions was rare, around 4 per cent, mostly note-taking or photographing the screen. The authors' conclusion is the practical one: supervision matters most for tasks that are easy to cheat, such as memory tasks, and less for complex reasoning in low-stakes conditions.

The honest reading is therefore narrower than "unproctored tests are inflated". It is that the effect of supervision depends on the task and the stakes, and that you cannot know which applies to a random web test because nothing about it is documented.